Doctors’ Migration in 2018
Ruri Proto
05/06/2021
Research Questions
To which geographical regions do foreignly trained doctors migrate?
From which geographical regions do foreignly trained doctors emigrate?
Are there any distinguishable patterns? For example:
- do doctors from a certain region emigrate to a select few other regions?
- are there any regions to which doctors do not migrate?
Data Origin
I retrieved data from the OECD website, where I downloaded migration information from all OECD countries in 2018. I subsequently retrieved a dataset linking countries with geographical subregions from github.
library(readr)
library(here)
wf_mig = read_csv(here("data-raw/workforce-migration.csv")) #migration data
sub_reg = read_csv(here("data-raw/subregions.csv")) #subregions dataThe workforce migration data set includes the number of foreignly trained doctors who, in 2018, were registered or in the process of gaining registration to practise in a country other than the one in which they had obtained their medical education qualifications - this includes medical interns and residents.
head(wf_mig)## # A tibble: 6 x 11
## COU Country VAR Variable CO2 `Country of ori~ YEA Year Value
## <chr> <chr> <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl>
## 1 CAN Canada FTDS Foreign-traine~ AFG Afghanistan 2018 2018 5
## 2 FRA France FTDS Foreign-traine~ AFG Afghanistan 2018 2018 12
## 3 DEU Germany FTDS Foreign-traine~ AFG Afghanistan 2018 2018 153
## 4 NZL New Zeal~ FTDS Foreign-traine~ AFG Afghanistan 2018 2018 0
## 5 NOR Norway FTDS Foreign-traine~ AFG Afghanistan 2018 2018 22
## 6 CHE Switzerl~ FTDS Foreign-traine~ AFG Afghanistan 2018 2018 6
## # ... with 2 more variables: Flag Codes <lgl>, Flags <lgl>The subregions dataset links countries with their respective geographical regions and subregions. I focused on subregions in my analysis because their number is more manageable than that of countries (far too many!) or regions (far too few!), which will enhance the readability of the plot.
head(sub_reg)## # A tibble: 6 x 11
## name `alpha-2` `alpha-3` `country-code` `iso_3166-2` region `sub-region`
## <chr> <chr> <chr> <chr> <chr> <chr> <chr>
## 1 Afghanis~ AF AFG 004 ISO 3166-2:~ Asia Southern Asia
## 2 Åland Is~ AX ALA 248 ISO 3166-2:~ Europe Northern Eur~
## 3 Albania AL ALB 008 ISO 3166-2:~ Europe Southern Eur~
## 4 Algeria DZ DZA 012 ISO 3166-2:~ Africa Northern Afr~
## 5 American~ AS ASM 016 ISO 3166-2:~ Ocean~ Polynesia
## 6 Andorra AD AND 020 ISO 3166-2:~ Europe Southern Eur~
## # ... with 4 more variables: intermediate-region <chr>, region-code <chr>,
## # sub-region-code <chr>, intermediate-region-code <chr>Data Processing
You can see the full script here for a detailed explanation of all data processing steps taken.
I wanted my final dataset to include subregions of origin, destination subregions, and number of migrants. This required me to join wf_mig and sub_reg by ISO Alpha-3 codes. These are standardised codes for countries, and thus unique identifiers that function as joining keys.
Removing unnecessary data
I removed:
- flows between identical countries (not interested in domestic migration)
- unnecessary columns, e.g., the year of migration (since data is solely from 2018) or information about regions (since I was only interested in subregions).
I also gave columns suggestive, shorter names.
library(dplyr)
#wf_mig - keep relevant columns and rename them
wf_mig = wf_mig[, c("COU", "Country", "CO2", "Country of origin", "Value")]
wf_mig = wf_mig %>%
rename(code_to = "COU",
country_to = "Country",
code_from = "CO2",
country_from = "Country of origin",
number = "Value"
)
#wf_mig - remove domestic migration
wf_mig = wf_mig %>%
filter(country_to != country_from)
#sub_reg - keep relevant columns and rename them
sub_reg = sub_reg[, c("name", "alpha-3", "sub-region")]
sub_reg = sub_reg %>%
rename(country = name,
code = "alpha-3",
subregion = "sub-region")Joining data frames
I allocated subregions to each country by joining wf_mig and sub_reg by country code. I also provided helpful names to distinguish between origin and destination subregions.
#join datasets based on country code
data = left_join(wf_mig, sub_reg,
by = c("code_to" = "code")
) #add subregions for destination countries
data = rename(data, subregion_to = subregion) #destination subregion
data = left_join(data, sub_reg,
by = c("code_from" = "code")
) #add subregions to origin countries
data = rename(data, subregion_from = subregion) #subregion of originEvaluating the final dataset
The final dataset had no missing values.
sapply(data,
function(x) sum(is.na(x))
)## code_to country_to code_from country_from number
## 0 0 0 0 0
## subregion_to subregion_from
## 0 0head(data)## # A tibble: 6 x 7
## code_to country_to code_from country_from number subregion_to subregion_from
## <chr> <chr> <chr> <chr> <dbl> <chr> <chr>
## 1 CAN Canada AFG Afghanistan 5 Northern Ame~ Southern Asia
## 2 FRA France AFG Afghanistan 12 Western Euro~ Southern Asia
## 3 DEU Germany AFG Afghanistan 153 Western Euro~ Southern Asia
## 4 NZL New Zealand AFG Afghanistan 0 Australia an~ Southern Asia
## 5 NOR Norway AFG Afghanistan 22 Northern Eur~ Southern Asia
## 6 CHE Switzerland AFG Afghanistan 6 Western Euro~ Southern AsiaData Transformations
The code for this work is here.
The visualization I planned followed the procedure of Sander et al. (2014) and required two objects:
flow_matrix- a matrix containing the number of migrants between all combinations of subregions (origin subregions as rows, destination subregions as columns)subregion_details- a data frame showing plotting parameters (e.g., colour codes) for each subregion.
The flow matrix
I created the subregions data frame, containing the total number of migrants per subregion of origin and destination, irrespective of country:
library(dplyr)
library(reshape2)
#get number of immigrants/emigrants at subregion level
subregions = data %>%
group_by(subregion_from, subregion_to) %>%
summarize(subregion_number = sum(number))
#convert subregions data frame into wide format
subregions = dcast(subregions,
subregion_from ~ subregion_to, #origin subregions as rows
value.var = "subregion_number"
) #convert into wide format
#set rownames to subregion_from to facilitate indexing
rownames(subregions) = subregions$subregion_from
head(subregions)## subregion_from
## Australia and New Zealand Australia and New Zealand
## Central Asia Central Asia
## Eastern Asia Eastern Asia
## Eastern Europe Eastern Europe
## Latin America and the Caribbean Latin America and the Caribbean
## Melanesia Melanesia
## Australia and New Zealand Eastern Europe
## Australia and New Zealand 2689 1
## Central Asia 3 95
## Eastern Asia 756 32
## Eastern Europe 182 7343
## Latin America and the Caribbean 69 16
## Melanesia 67 NA
## Latin America and the Caribbean
## Australia and New Zealand 2
## Central Asia 2
## Eastern Asia 1
## Eastern Europe 113
## Latin America and the Caribbean 10674
## Melanesia NA
## Northern America Northern Europe
## Australia and New Zealand 697 941
## Central Asia 26 95
## Eastern Asia 527 336
## Eastern Europe 1699 11632
## Latin America and the Caribbean 2064 900
## Melanesia 4 5
## Southern Europe Western Asia Western Europe
## Australia and New Zealand 1 85 44
## Central Asia 2 69 442
## Eastern Asia NA NA 539
## Eastern Europe 92 10892 24244
## Latin America and the Caribbean 6 652 2159
## Melanesia NA NA 1I initialized a flow matrix with all subregions as rows and columns which contained only 0s, treating rows as origin subregions and columns as destination subregions.
I updated the values in the flow_matrix with the ones in subregions. This approach ensured that all possible combinations of subregions were present in flow_matrix, even if they were absent in the subregions data set. Absence would indicate those combinations of subregions had 0 migration levels.
#find all subregions in the dataset
unique_subreg = unique(c(unique(data$subregion_to), unique(data$subregion_from)))
#update flow_matrix with values from subregions
for(i in unique_subreg) { #take each unique subregion
for(j in unique_subreg) { #combine it with all subregions
flow_matrix[i, j] = ifelse( #for each combination
(flow_matrix[i, j] != subregions[i, j] && #if subregions value is different from flow_matrix value
!(is.na(subregions[i, j]))), #providing subregions value is not missing
subregions[i, j], #replace value in flow_matrix with subregions value
flow_matrix[i, j] #otherwise keep 0 in flow_matrix
)
}
}At the end, the flow matrix looked like this:
head(flow_matrix)## Eastern Asia South-Eastern Asia Sub-Saharan Africa
## Eastern Asia 0 0 0
## South-Eastern Asia 0 0 0
## Sub-Saharan Africa 0 0 0
## Northern Africa 0 0 0
## Southern Europe 0 0 0
## Northern America 0 0 0
## Northern Africa Southern Europe Northern America
## Eastern Asia 0 0 527
## South-Eastern Asia 0 0 488
## Sub-Saharan Africa 0 0 3849
## Northern Africa 0 0 1895
## Southern Europe 0 956 459
## Northern America 0 1 1218
## Latin America and the Caribbean Western Asia
## Eastern Asia 1 0
## South-Eastern Asia 1 2
## Sub-Saharan Africa 1 125
## Northern Africa 0 228
## Southern Europe 128 1570
## Northern America 8 534
## Australia and New Zealand Southern Asia Eastern Europe
## Eastern Asia 756 0 32
## South-Eastern Asia 1086 0 4
## Sub-Saharan Africa 2828 0 2
## Northern Africa 67 0 12
## Southern Europe 94 0 110
## Northern America 784 0 7
## Northern Europe Western Europe
## Eastern Asia 336 539
## South-Eastern Asia 819 571
## Sub-Saharan Africa 5931 2112
## Northern Africa 5622 8771
## Southern Europe 4800 14367
## Northern America 172 463The subregion details data frame
This data frame included colours for circle sectors, circle links, and the total flow of migrants (immigrants + emigrants) in each country. Code for this section has been largely adapted from Sander et al. (2014).
I added the number of emigrants, immigrants, and total migrants for each subregion to the newly created subregion_details dataframe.
#Compute number of emigrants per subregion
df_from = data %>%
group_by(subregion_from) %>%
summarize(emig = sum(number))
#Compute number of immigrants per subregion
df_to = data %>%
group_by(subregion_to) %>%
summarize(immig = sum(number))
#create subregion_details data frame with info about total migration flow
subregion_details = left_join(df_from,
df_to,
by = c("subregion_from" = "subregion_to")
)
##I am aware I could have done this by summing the rows and columns of `flow_matrix`
##but I wanted to do things this way so that I could compare the two outputs
##and hopefully find they are identical as a way to check my work (they were!). Because circular plots offer limited space, I wanted to eliminate subregions with few migrants from the dataset, but also give the user the choice to include as many subregions as they want. In my case, I excluded subregions that had the bottom 20% number of total migrants.
#eliminate subregions where the total number of migrants is below the given quantile
(tiny_subreg = subset(subregion_details,
total < quantile(total, 0.2) #user-defined quantile
))
#remove tiny subregions from subregion_details
subregion_details = subregion_details[!(subregion_details$subregion %in% tiny_subreg$subregion), ]I then assigned colours to each available subregion after ordering them ascendently by the total number of migrants. This process runs automatically independently of how many subregions out of the total of 17 the user selects.
#add rgb codes to each subregion
rgb_pool = c("255,0,0", #red
"0,255,0", #lime
"128,128,0", #olive
"148,0,211", #dark violet
"0,206,209", #dark turquoise
"255,0,255", #magenta
"128,0,0", #maroon
"255,99,71", #tomato
"0,128,0", #green
"0,0,255", #blue
"128,0,128", #purple
"0,128,128", #teal
"0,0,128", #navy
"250,128,144", #salmon
"100,149,237", #corn flower blue
"153,50,204", #dark orchid
"60,179,113" #medium sea green
) #googled 17 rgb codes that enhance contrast; 17 = length(unique_subreg)
#select as many colours as needed depending on the amount of subregions included
subregion_details$rgb = rgb_pool[1:nrow(subregion_details)]I then stored two versions of these colours in HEX format (one of which had increased transparency) for different elements of the graph. To accomplish this, I split colour codes into individual variables, and then used the rgb function to increase transparency and convert them into HEX codes.
#Split rgb codes into 3 variables
n = nrow(subregion_details)
subregion_details = cbind(subregion_details, #split codes and treat them as numbers
matrix(as.numeric(unlist(strsplit(subregion_details$rgb, split = ","))),
nrow = n, byrow = TRUE
) #arrange them in n columns in a matrix
)
subregion_details = subregion_details %>%
rename( #rename columns according to the colour index
r = '1',
g = '2',
b = '3',
)
#increase transparency and transform into HEX codes
subregion_details$rcol = rgb(subregion_details$r,
subregion_details$g,
subregion_details$b,
max = 255
)
subregion_details$lcol = rgb(subregion_details$r,
subregion_details$g,
subregion_details$b,
alpha = 200, #transparency index
max = 255
)I also ordered rows in subregion_details ascendently by the total number of migrants to facilitate indexing later on, and added the xmin = 0 and xmax = subregion_details$total columns which will demarcate axis limits in the plot (from 0 to total amount of migrants) for each subregion.
At the end, the subregion_details data frame looked like this:
head(subregion_details)## subregion emig immig total order rgb r g b rcol
## 1 Eastern Asia 2191 0 2191 5 255,0,0 255 0 0 #FF0000
## 2 South-Eastern Asia 2971 0 2971 6 0,255,0 0 255 0 #00FF00
## 3 Sub-Saharan Africa 14848 0 14848 7 128,128,0 128 128 0 #808000
## 4 Northern Africa 16595 0 16595 8 148,0,211 148 0 211 #9400D3
## 5 Southern Europe 22484 1085 23569 9 0,206,209 0 206 209 #00CED1
## 6 Northern America 3187 24303 27490 10 255,0,255 255 0 255 #FF00FF
## lcol xmin xmax
## 1 #FF0000C8 0 2191
## 2 #00FF00C8 0 2971
## 3 #808000C8 0 14848
## 4 #9400D3C8 0 16595
## 5 #00CED1C8 0 23569
## 6 #FF00FFC8 0 27490Data Visualization
The data visualization I have selected is a circular plot diagram and you can find the code here.
Background information
Essentially, this plot draws tracks on a circle and splits them into sectors to reflect differences between subregions in migrant numbers. It also plots links to illustrate the flow of migrants between subregions.
Readers who are interested in the inner workings of the code should consult Sander et al. (2014) and Gu (2020). I used both of these resources, but I prefer the latter because it offers an in-reasonable-depth explanation of the circlize package.
Preliminary circular plots
I started by setting some plotting parameters related to the size of the circle, padding, sector gaps, etc.
suppressPackageStartupMessages(library(circlize))
circos.clear() #reset circular layout parameters
par(mar = c(0, 0, 0, 0)) #margin around chart
circos.par(cell.padding = c(0, 0, 0, 0),
track.margin = c(0, 0.1),
start.degree = 45, #start plotting at 2 o'clock
gap.degree = 2, #gap between circle sectors
points.overflow.warning = FALSE,
canvas.xlim = c(-1.3, 1.3), #size of circle
canvas.ylim = c(-1.3, 1.3) #size of circle
)I then initialized the layout to allocate subregions into sectors whose sizes are bounded by xmin and xmax. This approach ensures the relative size of sectors matches the relative difference in migration for each subregion.
circos.initialize(factors = subregion_details$subregion, #allocate sectors on circle to subregions
xlim = cbind(subregion_details$xmin,
subregion_details$xmax) #set limits of the x axis for each sector between 0 and total = xmax
)The next step involved creating a plotting region to which I added graphics.
- I plotted subregion names using
circos.text() - I plotted a first track using
circos.rect()which was split into one sector per subregion reflecting the total amount of migrants (emigrants + immigrants).
This initial work resulted in:
circos.initialize(factors = subregion_details$subregion, #allocate sectors on circle to subregions
xlim = cbind(subregion_details$xmin,
subregion_details$xmax) #set limits of the x axis for each sector between 0 and total = xmax
)
options(scipen = 10) #prevent scientific notation on plot
circos.trackPlotRegion(ylim = c(0, 1), #y-axis limits for each sector
factors = subregion_details$subregion,
track.height = 0.1,
panel.fun = function(x, y) { #for each new cell (i.e., intersection between sector and track)
name = get.cell.meta.data("sector.index") #retrieve cell meta data
i = get.cell.meta.data("sector.numeric.index")
xlim = get.cell.meta.data("xlim")
ylim = get.cell.meta.data("ylim")
#plot subregion names
circos.text(x = mean(xlim), #position text at middle of sector
y = ifelse(str_length(name) > 25, 4.5,
ifelse(str_length(name) > 20, 4,
ifelse(str_length(name) >= 14, 3.3, 3))
), #distance from plot depending on length of character
labels = name, #name of subregion
facing = "clockwise",
niceFacing = TRUE,
cex = 0.5, #scale text
col = subregion_details$rcol[i]
)
#plot a sector for each subregion
circos.rect(xleft = xlim[1],
ybottom = ylim[1],
xright = xlim[2],
ytop = ylim[2],
col = subregion_details$rcol[i], #use less transparent colours
border = subregion_details$rcol[i]
)
}
)
This plot shows the relative number of total migrants (immigrants + emigrants) for each subregion. I next added another track where:
- the coloured portions represent relative numbers of emigrants per subregion
- the white portions represent relative numbers of immigrants per subregion.
This was achieved by including the following code inside the panel.fun() function in the previous code chunk.
#distinguish between immigrants and emigrants in each subregion
circos.rect(xleft = xlim[1],
ybottom = ylim[1],
xright = xlim[2] - rowSums(flow_matrix)[i], #i.e., total - emigrants
ytop = ylim[1] + 0.3,
col = "white",
border = "white"
)
#add a white contour to separate the previous two rectangles
circos.rect(xleft = xlim[1],
ybottom = 0.3,
xright = xlim[2],
ytop = 0.32,
col = "white",
border = "white"
)
Final plot
Next, I included links between origin subregions and destination subregions, in order to show migration patterns, and an axis to give an indication of actual migrant numbers. This required some further data processing to transform flow_matrix into its long format and add parameters to guide the position of links, sums1 and sums2.
#plot links for each combination of regions
for(k in 1:nrow(flow_matrix_long)){ #for each row in the flow matrix
i = match(flow_matrix_long$subregion_from[k],
subregion_details$subregion) #get plotting details for subregion of origin
j = match(flow_matrix_long$subregion_to[k],
subregion_details$subregion) #get plotting details for destination subregion
circos.link(sector.index1 = subregion_details$subregion[i], #need to identify indices to identify
point1 = c(subregion_details$sum1[i],
subregion_details$sum1[i] + abs(flow_matrix[i, j])), #starting point of link
sector.index2 = subregion_details$subregion[j],
point2 = c(subregion_details$sum2[j],
subregion_details$sum2[j] + abs(flow_matrix[i, j])), #endpoint of link
border = subregion_details$lcol[i],
col = subregion_details$lcol[i], #use the more transparent colour to increase visibility
)
#update sum1 and sum2 to move along the circle into the next sector
subregion_details$sum1[i] = subregion_details$sum1[i] + abs(flow_matrix[i, j])
subregion_details$sum2[j] = subregion_details$sum2[j] + abs(flow_matrix[i, j])
}As a result, I produced this plot - which is linked here as a png file because R output in this package looks horrendous.
The colour of links suggests their origin. Notice that immigration patterns are plotted through links starting in the emigrants section of a subregion (the second coloured arc inwards), and ending up in the immigrants section of another (or the same) subregion (the white arc continuing from the emigrant arc). The size of the link gives the relative amount of migrants moving from one subregion to another.
For example, the migration flow in Southern Asia consists exclusively of emigrants, most of whom tend to go to Northern Europe, with far fewer going to Australia and New Zealand or Northern America. As we can see, immigrants from Southern Asia make up a sizeable chunk of the total amount of immigrants in Northern Europe (the blue ribbon corresponds to about 25,000 immigrants in Northern Europe out of a total of about 75,000, as demarcated by the white stripe of the sector).
Summary
Findings show that:
Western Europe has the most migrants, closely followed by Northern Europe and Eastern Europe.
Whilst Western and Northern Europe have relatively low proportions of emigrants and high proportions of immigrants, this is completely the opposite in Eastern Europe.
Most doctors in Latin America and the Caribbean emigrate to countries within the same region.
Very few doctors emigrate to Eastern Europe.
Reflection
If I had more time to spend on this project:
I would fix the issue that causes some links to go ever so slightly beyond the borders for emigrants - yes, Northern Africa, I’m looking at you! I am still not sure what the reason for this is, but the
circlizepackage still receives updates, so I will keep an eye on it.I would figure out how to automatically adjust the distance between text and plot depending on the length of the string whilst avoiding hardcoding string length values.
I would extend this project beyond this plot to understand why healthcare workers migrate to specific subregions - Is it better pay? Is it the chance to work in specific areas of medicine?
I would evaluate how migration flow patterns change over time, and visualise this with an interactive plot. There are some nice integrations available with JavaScript that can make this possible.
I would try to replicate this plot with a recent addition to the
circlizepackage,chordDiagram().
Resources used
To complete this project, I used several resources:
Sander et al. (2014) - the minds behind adapting circular plots to explore migration flows
Gu (2020) - in-reasonable-depth resource for th
circlizepackagestackoverflow.com - I am gradeful to random skilled strangers for helping me learn new ways of doing things
Xie et al. (2021) - fantastic, in-depth resource for tackling RMarkdown issues
You can find my repository here.